Point-Cloud Completion with Pretrained Text-to-image Diffusion Models Gal Chechik NVIDIA Research 2